// ContextLeak in AI Agents // The whole attack surface here is a tool name and a tool description. Stealing an LLM agent's runtime context, meaning the user prompt, the execution trajectory and the tool list, needs three things to line up. The agent has to pick the malicious tool, it has to pass its context in as arguments, and the tool has to forward that anywhere the attacker wants. Existing work covers the first and third conditions and leaves the second one mostly alone. ContextLeak targets the middle step. Researchers at Duke use an attack LLM to generate the malicious tool's name and description, then fine-tune that LLM with reinforcement learning on a set of shadow users with diverse simulated agent contexts. The reward functions are built specifically for the exfiltration objective. It remains highly effective when the shadow contexts differ substantially from the victim's, and it outperforms existing malicious-tool attacks adapted to this setting. Paper: Chat with Paper:

Context Corner 2.0 is here! 🍽️🤖 Loved the first conversation on context, memory, and agents? We’re back for round two, with even more time around the table and an incredible lunch to match! If you're a developer building, experimenting, or stuck on the frontlines of AI memory and agents, come hungry and ready to chat. 📍 San Francisco, CA 🗓️ Friday, Aug 14 | 12:00 PM - 1:30 PM PDT Here's the registration link:

🆕 Context Engineering in 2026: Compaction, Memory & Cost @Whats_AI, @samridhivaid and @omar_solano1 return! This workshop is about engineering the context window so rot stops happening, shown with @towards_AI's open-source AI tutor, which answers questions for students of our AI-engineering courses. Context engineering is deciding what the model sees on every single call — instructions, history, retrieved course content, memory, and tool outputs — and it's the line between a tutor that holds a coherent session and one that forgets the student's setup halfway through. We'll move in three stages, mirroring how the project actually went. The concepts: - the two root problems (a finite window, a stateless model), - the full compaction toolkit (truncation, trimming, tool-result clearing, summarization, and offloading to files — and when each actually helps), - memory that survives across sessions, skills loaded on demand, and - production-grade retrieval (chunking, metadata, course scoping, hybrid search, reranking, and evaluating). We'll cover the tutor's architecture, and the evaluation harness we used to measure every run on Gemini — tokens, cost, latency, and memory probes instead of vibe-checks. At real volume, even Gemini Flash got expensive, so we tested whether open and local models could match the quality for a fraction of the cost and match result quality. Everything is open-source and will be shared during the workshop.